AutoNovel: Automatically Discovering and Learning Novel Visual Categories.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Discovering novel classes in image collections with limited labelled data is challenging.
• Existing methods often introduce bias by training image representations solely on labelled data.

Purpose of the Study:

• To present AutoNovel, a novel approach for discovering new image classes from labelled and unlabelled data.
• To improve image representation learning by avoiding bias and enhancing clustering capabilities.

Main Methods:

• Utilizing self-supervised learning on combined labelled and unlabelled data to train image representations from scratch.
• Employing ranking statistics to transfer knowledge from labelled to unlabelled data for clustering.
• Optimizing a joint objective function for improved supervised classification and unsupervised clustering.

Main Results:

• AutoNovel substantially outperforms current methods in novel category discovery benchmarks.
• The approach effectively clusters unlabelled images, demonstrating strong performance in fully unsupervised image clustering.
• A method for estimating the number of unknown classes a priori was proposed and evaluated.

Conclusions:

• AutoNovel offers a robust solution for novel class discovery by leveraging self-supervised learning and knowledge transfer.
• The method enhances both supervised classification and unsupervised clustering tasks.
• AutoNovel shows significant potential for various image analysis applications, including unsupervised clustering.